GSA Annual Meeting in Phoenix, Arizona, USA - 2019

Paper No. 143-14
Presentation Time: 5:00 PM


GUNDERSON, Kellen L.1, ZHANG, Zoe1 and CHENG, Shuxing2, (1)Chevron Energy Technology Company, 1500 Louisiana St., Houston, TX 77002, (2)Chevron Information Technology Company, 1500 Louisiana St., Houston, TX 77002

With the recent emergence of machine learning technologies in many industries, there is an immense interest in developing machine learning applications to applied geoscience problems. This has highlighted the importance of geoscientists understanding the capabilities and limitations of machine learning technologies and developing appropriate use cases for their disciplines. Here, we outline the philosophical underpinning, current progress, and potential pitfalls of machine learning to applied structural geology problems, with a focus on the interpretation and characterization of subsurface structures for hydrocarbon exploration and development. The types of problems petroleum structural geologists solve include building 2D/3D subsurface structural models from wells and seismic data (geometry), reconstructing the evolution of structures through time (kinematics), and determining the underlying causes of the geometric and kinematic models (dynamics). Machine learning, built upon statistical theory and computer science, aims to derive mathematical models from training data to perform specific tasks without explicit supervision. It encompasses a wide suite of models and algorithms that can be applied to a subset of petroleum structural geology problems. An example is the use of computer vision and deep learning to aid in seismic structural interpretation of faults and salt bodies. In these use cases, a variety of deep learning approaches have been studied, such as: multiclass classification, semantic segmentation of the seismic images, analyzing synthetic and real data sets, and comparing analyses on the seismic data and image attributes extracted from the seismic. As the applications of machine learning in petroleum structural geology increase, potential pitfalls need to be avoided, including building proper training datasets, introducing interpretability to the deep learning models, building physics-based machine learning models to integrate geological priors, adapting to large scale data volumes, and pushing 2D machine learning solutions into 3D. Through experimentation with these new techniques and by addressing potential pitfalls, applied structural geologists can open a new area of research that can further our fundamental understanding of earth structures.